Can. J. Remote Sensing, Vol. 31, No. 5, pp. 400–411, 2005

Variability of seasonal CASI image data productsand potential application for management zone

delineation for precision agricultureJiangui Liu, John R. Miller, Driss Haboudane, Elizabeth Pattey, and Michel C. Nolin

Abstract. The delineation of management zones is an important step to implementing site-specific crop managementpractices. Remote sensing is a cost-effective way to acquire information needed for delineating management zones, since ithas been successfully used for mapping soil properties and monitoring crop growth conditions. Remotely sensedhyperspectral data are particularly effective in deriving crop biophysical parameters in agricultural fields; therefore, thepotential of hyperspectral data to contribute to management zone delineation needs to be assessed. In this study, the spatialvariability of soil and crops in two agricultural fields was studied using seasonal compact airborne spectrographic imager(CASI) hyperspectral images. Different spectral features including soil brightness and colouration indices, principalcomponents of soil reflectance data, and crop descriptors (leaf area index (LAI) and leaf chlorophyll content) were derivedfrom CASI data and used to partition the fields into homogeneous zones using the fuzzy k means unsupervised classificationmethod. The reduction of variances of soil electrical conductivity, LAI, leaf chlorophyll content, and yield was inspected todetermine the appropriate number of zones for each field. The zones obtained were interpreted according to the soil surveymap and field practices. Analysis of variance (ANOVA) was conducted to examine the effectiveness of the delineation. Thestudy shows that the spatial patterns of the resulting soil zones faithfully represent the soil classes described by the soilsurvey maps, and the spatial patterns of the resulting crop classes discriminated the different crop growth conditions well.These results show that hyperspectral data provide important information on field variability for management zonedelineation in precision agriculture.

Résumé. La délimitation des zones de gestion homogènes est une étape importante dans la mise en place des procédures degestion localisée des ressources agricoles. La télédétection peut s’avérer éonomiquement viable pour l’acquisition desdonnées requises à la délimitation de ces zones. En effet, elle a déjà permis de cartographier des propriétés de sols et desuivre la croissance des cultures. Les données hyperspectrales sont très utiles pour dériver des descripteurs biophysiques deschamps en cultures; il faut donc évaluer le potentiel de la télédétection hyperspectrale à définir adéquatement la délimitationdes zones de gestion homogènes. À l’aide d’une série temporelle d’mages hyperspectrales du capteur aéroporté CASI(« compact airborne spectrographic imager »), la variabilité spatiale des propriétés du sol et des cultures dans deux champsagricoles ont été étudiés. Divers indicateurs spectraux, dont les indices de brillance et de coloration du sol, des composantesprincipales de réflectance du sol et des descripteurs du couvert végétal agricole (l’indice de surface foliaire (LAI) et lateneur en chlorophylle) ont été extraits des données CASI et utilisés pour segmenter les champs en zones homogènes àl’aide d’une classification non dirigée utilisant la méthode de groupement flou à k moyens. L’observation de la réduction dela variance de la conductivité électrique du sol, du LAI, de la teneur en chlorophylle des feuilles, et du rendement agricole apermis de déterminer le nombre approprié de zones homogènes dans chaque champ. Les résultats ainsi obtenus ont étéévalués et interprétès grâce à l’utilisation de la carte pédologique et des informations sur les pratiques agricoles. Uneanalyse de variance (ANOVA) a été réalisée pour évaluer la précision de la segmentation retenue. Les vérifications ontconfirmé que les zones homogènes déterminées à partir des propriétés spectrales du sol représentaient bien les classesdécrites sur la carte pédologique, et que les zones homogènes établies à partir des descripteurs biophysiques du couvertagricole décrivaient bien les diverses conditions de croissance des cultures étudiées. Cela montre bien que la télédétectionhyperspectrale est une source d’information importante pour la détection de la variabilité spatiale des champs agricoles ainsique pour la délimitation des zones de gestion homogènes en agriculture de précision.

Liu et al.411

400 © 2005 Government of Canada

Can. J. Remote Sensing, Vol. 31, No. 5, pp. 400–411, 2005

Received 15 December 2004. Accepted 10 June 2005.

J. Liu1 and E. Pattey. Agriculture and Agri-Food Canada, 960 Carling Avenue, Ottawa, ON K1A 0C6, Canada.

J.R. Miller. Department of Physics and Astronomy, York University, Toronto, ON M3J 1P3, Canada.

D. Haboudane. Département des Sciences Humaines, UQAC, Chicoutimi, QC G7H 2B1, Canada.

M.C. Nolin. Agriculture and Agri-Food Canada, Québec, QC G1W 2L4, Canada.

1Corresponding author (e-mail: [email protected]).

IntroductionOne of the important inputs to site-specific management

practices in agriculture is the delineation of management zones.A management zone is defined as a portion of a field thatexpresses a homogeneous combination of yield-limiting factorsfor which a single rate of a specific crop input is appropriate(Doerge, 1998). The delineation of management zones relies onthe exploitation of spatial variability of the agriculture field.Zhang et al. (2002) classified the variability into six groups:yield variability, field variability, soil variability, cropvariability, variability in anomalous factors, and managementvariability. Information on the variability can be ascribed asfollows: (i) seasonally stable conditions, such as yield-based orsoil-based management units, which need to be determinedonly once every season; and (ii) seasonally variable conditions,such as soil moisture, weeds, and crop disease, which need tobe monitored continuously during the season (Moran et al.,1997).

Remote sensing offers a quick and cost-effective way toobtain information on the variability of agricultural fields, suchas soil properties, crop vigour, crop stress, and relative cropyield (Moran et al., 1997). Remotely sensed hyperspectral datahave been successfully used in crop studies for estimation ofbiophysical descriptors (Haboudane et al., 2002; 2004;Thenkabail et al., 2000), prediction of crop vigour and yield(Tomer et al., 1995; Shibayama and Akiyama, 1991), andmonitoring of environmental impact (Strachan et al., 2002;Pattey et al., 2001; Leone and Escadafal, 2001; Lelong et al.,1998). These studies demonstrated that hyperspectral remotesensing provides a powerful tool for precision agricultureapplications.

The objective of this study was to explore the potential andability of hyperspectral remote sensing data for managementzone delineation in precision agriculture. Crop fields weredelineated into homogeneous zones using soil and cropproperties extracted from multitemporal compact airbornespectrographic imager (CASI) hyperspectral data, and theacquired zones were interpreted according to the soil surveymaps and the treatments applied in the fields.

Study site and hyperspectral dataThe study site is located in the former greenbelt farm of

Agriculture and Agri-Food Canada, Ottawa, Ontario, Canada(45°18′N, 75°45′W). The two neighbouring fields investigatedin the present study are referred to as fields 25 and 23. Field 25is primarily composed of two soil associations (D3, Brandonseries; M3–NG2, Montain, Allendale, and North Gower series)that share similar drainage conditions (poorly drained) andtaxonomic classification (orthic humic gleysol). They aredifferentiated by the subsurface texture, which is finer in D3(silty clay loam to clay loam) than in M3–NG2 (sandy clayloam to fine sandy loam). Field 23 is composed of seven soillandscape units with variable drainage classes, profile textures,and genetic evolution (Perron et al., 2002). Figure 1 shows thedetailed soil survey map of the two fields, and Table 1 gives thesoil classification legend.

A survey was made in the two fields in November 2002(Perron et al., 2003) to obtain soil electrical conductivity at twodepths, namely 0–30 cm (EC30) and 0–100 cm (EC100). In theyear 2001, uniform nitrogen (N) was applied in field 23, and aspecific N application pattern was imposed on field 25 (seeFigure 5d later in the paper).Yield data were acquired during

© 2005 Government of Canada 401

Canadian Journal of Remote Sensing / Journal canadien de télédétection

Figure 1. Detailed soil survey map (Marshall et al., 1979) of field 23 (upper half of map) andfield 25 (lower half of map). The description of soil landscape unit (SLU) is presented inTable 1.

harvest using a combine equipped with a yield monitor for bothfields. CASI hyperspectral data were collected four times in2000 and three times in 2001, spanning crop growingconditions, by intensive field campaigns (IFCs). CASI wasoperated in the hyperspectral mode with 2 m spatial resolutionand 7.5 nm bandwidth. The 72 spectral channels acquired bythe sensor covered the visible and near-infrared portions of thesolar spectrum, ranging from 408 to 947 nm. The data acquiredon 20 June 2000 were chosen for soil partitioning, as the twofields were almost bare of vegetation at that time. In 2001, cornand spring wheat were planted in fields 23 and 25, respectively.Acquisition dates in 2001 were planned to coincide withdifferent phenological development stages, providing imagedata covering the early, active growth and reproductive cropgrowth stages. The data from the three IFCs in 2001, acquiredon 14 June (IFC1), 26 June (IFC2), and 19 July (IFC3), wereused for crop field partitioning to study the spatial and temporalvariability of the two crop fields. CASI data were processed toabsolute ground reflectance by an operational processingprocedure, which includes radiance calibration, atmosphericcorrection using the CAM5S model, and flat field correction, asdescribed by Haboudane et al. (2004).

MethodsFeature extraction

Feature extraction and selection is a necessary step inhyperspectral data processing due to the large number ofspectral channels available. Effective methods for featureextraction are objective oriented. This can be demonstrated byrecently developed vegetation indices. The modified triangularvegetation index (MTVI2) is presented as an excellentestimator of leaf area index (LAI) that minimizes leafchlorophyll content variation (Haboudane et al., 2004),whereas the combined use of the transformed chlorophyllabsorption in reflectance index (TCARI) and the optimizedsoil-adjusted vegetation index (OSAVI) provides a good

estimation of leaf chlorophyll content while minimizing LAIvariation (Haboudane et al., 2002). Nevertheless, featureselection and extraction inevitably results in information loss;therefore, special care should be taken when implementing anyprocedure of feature extraction.

Soil reflectance has direct relationships with soil opticalproperties (colour and brightness) and other soil propertiessuch as texture, soil moisture, and organic matter content(Mattikalli, 1997). Soil brightness and colour are important indifferentiating between soil types (Leone and Escadafal, 2001).They are believed to be determined by the amount and state ofiron and the content of soil organic matter, carbonate, moisture,etc. Indeed, Huete and Escadafal (1991) concluded thatreflectance intensity (or brightness) represents the dominant orprincipal source of spectral variance among soils, whereas thedifference of spectral curve shape (chromatic) is secondary. Acommon practice to obtain brightness and chromaticinformation is to convert from a red, green, and blue (RGB)colour composite constructed with multispectral bands to a hue,saturation, and intensity (HSI) colour representation system. Inthe HSI system, the intensity (I) component representsbrightness information, and the hue (H) and saturation (S)components represent chromatic information. In this study, theI and S components are extracted from CASI soil reflectancedata of 2000 and are referred to as brightness index (BI) andcolouration index (CI). The formulae, presented by Liu andMoore (1990) and modified by Escadafal et al. (1994), are asfollows:

BI 800 670 550= + +( )/R R R 3 (1)

CI 800 550 800= −( )/R R R (2)

where R is the reflectance of the channel, with the centralwavelength (in nm) indicated by the subscript. BI is equivalentto the average reflectance of the three channels and is a measureof the brightness of the soil. CI is equivalent to a measure of the


Vol. 31, No. 5, October/octobre 2005

Parent material SLUa Slope (%) Soil series Soil taxonomy Drainage

Fine-textured marine material (40%–60% clay) D3 2–5 Brandon Orthic humic gleysol Poorly drained

Strongly acid, sandy veneer (25–100 cm) overclayey material

M3 1–3 Mountain Gleyed sombric brunisol Imperfectly drainedAllendale Orthic humic gleysol Poorly drained

M5 0.5–2.0 Allendale Orthic humic gleysol Poorly drainedMontain Gleyed sombric brunisol Imperfectly drained

M6 0–2 Allendale Orthic humic gleysol Poorly drained

Moderately fine textured marine material(25%–40% clay)

NG2 0–2 North Gower Orthic humic gleysol Poorly drained

Medium- to fine-grained deep sandy material(>100 cm)

U1 2–7 Carlsbad Orthic sombric brunisol Well drainedU2 2–5 Carlsbad Orthic sombric brunisol Well drained

Ramsayville Gleyed sombric brunisol Imperfectly drainedU7 1–2 Ramsayville Gleyed sombric brunisol Imperfectly drained

St. Samuel Orthic humic gleysol Poorly drainedaSoil landscape unit.

Table 1. Soil classification legend for the two studied fields (see Figure 1).

slope of the soil spectrum and therefore soil colour (Escadafalet al., 1994). Thus, BI and CI calculated using these twoformulae are the first features to be used for soil-basedpartitioning.

Principal component (PC) analysis is an effective way offeature extraction. It compresses information into a fewcomponents and is a powerful tool for feature reduction inhyperspectral data processing. Principal componenttransformation based on the covariance matrix of soilreflectance data of 2000 was applied to images of fields 23 and25. The first three components (PC1, PC2, PC3) made up99.5% of the spectral information in field 23 (82.5%, 16.2%,and 0.9% for the first, second, and third principal components,respectively) and 99.7% in field 25 (91.7%, 7.8%, and 0.2% forthe first, second, and third principal components, respectively).They accounted for almost the total variability of soilreflectance data, and thus they were used as another feature setfor soil-based partitioning for comparison with the soil BI andCI measures.

LAI and leaf chlorophyll content are two important cropdescriptors. They are critical to understanding biophysicalprocesses and for predicting growth and productivity (Tucker etal., 1980; Moran et al., 1997). Therefore, CASI multitemporalproducts of LAI and leaf chlorophyll content were used forcrop-based partitioning. The formulae for LAI estimation are asfollows (Haboudane et al., 2004):

MTVI21.5 1.2 2.5800 550 670 550

800

= − − −+

[ ( ) ( )]

( )

R R R R

R2 1 2 − − −( )6 5R R800 670 0.5(3)

LAI 0.2227 3.6566 MTVI2= ×exp( ) (4)

The formulae for leaf chlorophyll content estimation are asfollows (Haboudane et al., 2002):

OSAVI = 1.16(R800 – R670)/(R800 + R670 + 0.16) (5)

TCARI = 3[R700 – R670 – 0.2(R700 – R550)R700/R670] (6)

Chl = –33.3 ln(TCARI/OSAVI) – 19.7 (7)

where Chl represents leaf chlorophyll content (µg·cm–2). Thedata from the three IFCs were clustered in an attempt to revealthe crop spatial patterns and their temporal variation.

Overall, five sets of features were derived from CASIreflectance data and used for field partitioning: soil features BIand CI, soil features PCs (PC1, PC2, and PC3) from soilreflectance data of 2000, and crop features LAI and leafchlorophyll content derived from CASI crop reflectance datafor the three IFCs in 2001. The soil features represent therelatively stable properties of the field, whereas crop features ofthe three IFCs reveal the seasonally variable conditions in thefields.

Feature preprocessing

More than one feature is used in this study for fieldpartitioning to integrate different aspects of information.Although all the features were extracted from the same sourceof data, their typical dynamic ranges are quite different. Thevalues of the features within the range relative to 1%–99% ofthe cumulative histogram were scaled to [0, 1] through a linearstretch. One of the reasons for this processing is that the relativeimportance of the features to the delineation is unknown, andtherefore they were given the same weight via data stretching.This processing also eliminates the outliers from the typicaldistribution range.

Clustering method

Because the number of management zones and their spatialdistribution are unknown, unsupervised methods were used tocluster the field into homogeneous regions by dividing thefeature space. The features from the sites are then extracted andrelated to the measured variables at the same sites to define theclass map of the variable of interest. Since the proposition byBezdek (1981), fuzzy k means has become one of the mostwidely used unsupervised classification methods. It is the mostaccurate among the unsupervised methods to reproduce theground data in a complex landscape (Duda and Canty, 2002)and has been used by many researchers to classify remotelysensed image data. The FUZCLUS module provided in the PCIsoftware (PCI Geomatics Enterprises Inc. 2001) was used inour study to partition the selected features.

Determination of the number of zones

The number of management zones is determined by the sizeof the field, the natural variability within the field, and certainmanagement factors (Zhang et al., 2002). The choice of anappropriate number of classes is a prerequisite beforeperforming unsupervised classification. We determine theoptimum number of zones using the method used by Fridgen etal. (2000), which is based on the inspection of the relative totalwithin-class variance (RTWCV) reduction of selected fieldvariables:

RTWCVclass field

= − −∈= ∈∑∑ ∑[ ] / [ ]x xij

j Ci

C

i jj1

2 2µ µ (8)

where C is the number of zones, xij is an observation of thevariable from zone i, xj is an observation of the variable in thewhole field, µ i is the average of the variable in zone i, and µ isthe average in the whole field. As the number of zonesincreases, RTWCV will decrease and then level off. The valueat which RTWCV levels off, or stops decreasing significantly,is a reasonable estimate of the number of zones that can be usedto partition the field.

In this study, to determine the appropriate number of zones,fields 25 and 23 were partitioned into 2–7 zones using thederived soil and crop feature sets. For each of the partitioned



results, RTWCV was calculated for the selected variables. Theselected variables included (i) yield, which is often consideredas the ultimate dependent variable; (ii) LAI and leafchlorophyll content, which are the most important cropdescriptors; and (iii) soil electrical conductivity. Electricalconductivity was measured at depths of 0–0.3 and 0–1.0 m(Perron et al., 2003). The RTWCV values of these selectedvariables are plotted against the number of zones, and theappropriate number of zones was determined from the plots.

Another factor that should be taken into consideration is thespatial distribution of the samples in a given zone. Pixel-basedimage classification usually divides the feature space. Thus,pixels in a zone are continuous in feature space but are notnecessarily so in the spatial domain. From the point of view ofthe agricultural producer, management zones shouldencompass significant areas with continuous spatialdistribution. Postclassification spatial filtering improves thedelineation by removing the isolated small clusters. In thisstudy, the isolated clusters with fewer than the given number ofpixels (i.e., 16 in this study) were detected and marked. Foreach pixel in the marked clusters, its class attribute wasdetermined by inspecting its neighbour pixels: it was assignedto the class that appeared most in this neighbourhood.

Analysis of variance

Analysis of variance (ANOVA) was conducted to test thedifference among the delineated zones for the selected soil andcrop properties. The technique is a single-factor ANOVA, withthe zone identification as the independent variable and the fielddescriptors, such as yield, electrical conductivity, LAI, and leafchlorophyll content, as dependent variables. Rafter et al. (2002)concluded that Tukey’s test is the most useful for all pairwisecomparisons, and the actual family-wise error rate (FWER)exactly equals the specified value. Therefore, Tukey’s multiplecomparison method (MCM) was applied to test the differencebetween the means for the dependent variables in the delineatedzones.

Results and discussionDetermination of appropriate number of zones

Figures 2 and 3 show the variance reduction of the selectedvariables in fields 25 and 23, respectively. Results from fivedelineations are given: two soil delineations using BI, CI, andprincipal components (PCs) and three crop delineations usingLAI and leaf chlorophyll content at IFC1, IFC2, and IFC3.Variance reduction of all the variables is given for the soil



Figure 2. Variance reduction by partitioning field 25 into 2–7 classes using (a) soil brightness and colouration indices(BI and CI, respectively), (b) the first three principal components (PCs) of soil reflectance, and (c–e) leaf area index(LAI) and chlorophyll (Chl) content from IFC1, IFC2, and IFC3, respectively. RTWCV, relative total within-classvariance.

delineations, and variance reduction of yield, electricalconductivity, and LAI and leaf chlorophyll content at thespecific IFC is given for the crop delineations. In Figures 2 and3, EC30 and EC100 refer to electrical conductivity between 0and 0.3 m and 0 and 1.0 m depth; LAI1, LAI2, and LAI3 andChl1, Chl2, and Chl3 represent LAI and leaf chlorophyllcontent at IFC1, IFC2, and IFC3, respectively.

Three to four zones were recommended for field 25 from aninspection of Figure 2. Based on BI and CI, classification offield 25 into four soil zones reduces the variances of EC30,EC100, LAI1, LAI2, and Chl1 to 58%, 65%, 54%, 78%, and78%, respectively. The results using principal components werealmost the same for the first three descriptors, with thevariances of the variables specified previously reduced to 59%,66%, and 71%, and with a limited reduced variance of LAI2and Chl1 to 93% and 91%, respectively. The soil features asidentified by hyperspectral reflectance seem to appropriatelyreveal the soil properties, as indicated by variance reduction ofsoil electrical conductivity. Inherent soil fertility indicators likesoil texture components (sand, silt, and clay content) andexchangeable cations (Ca and Mg) are closely related to soilelectrical conductivity (Nolin et al., 2002; Perron et al., 2002).Soil properties highly influenced by soil fertility managementlike soil pH and soil tests (available P and K), however, are lessclosely related to soil electrical conductivity (Perron et al.,

2002; 2003). BI–CI and principal components classificationssignificantly reduced the variances of yield and Chl3 to about85%, which tends to indicate that the detected soil propertieshad a restricted impact on growth conditions toward the end ofthe growing season in this field. In this field, soil propertiesseemed to explain mainly the variability related to theemergence of the spring wheat. With the progression of thegrowing season, soil properties captured by soil features did notsignificantly impact the variability of LAI and leaf chlorophyllcontent, indicating that there was no detection of N limitation.

Based on LAI and leaf chlorophyll content, classification offield 25 into four crop zones reduces the variances of LAI toabout 23%, 16%, and 34% and those of leaf chlorophyll contentto 37%, 54%, and 24% at IFC1, IFC2, and IFC3 stages,respectively. The crop zones delineated at IFC1 reduced thevariances of EC30 and EC100 to 77% and 82%, respectively.This also indicates that the crop growth condition revealed byLAI and leaf chlorophyll content at IFC1 stage is more affectedby the soil properties than those at IFC2 and IFC3. Anotherobservation is that the crop zones delineated at IFC1 and IFC2reduced the variance of yield to about 82% and 85%,respectively. Thus crop descriptors LAI and leaf chlorophyllcontent at the earlier development stages have more impact onwheat yield in field 25 than later on.



Figure 3. Variance reduction by partitioning field 23 into 2–7 classes using (a) soil brightness and colouration indices(BI and CI, respectively), (b) the first three principal components (PCs) of soil reflectance, and (c–e) LAI andchlorophyll.

Figure 3 suggests that two to three zones are recommendedfor field 23. When the field was partitioned into three soilzones, the variances of EC30 and EC100 were reduced to about64% and 73% when BI and CI were used and to about 76% and81% when PCs were used. Again, the soil features extractedfrom hyperspectral data seemed to capture the variability ofsome soil properties in the field. The variances of LAI and leafchlorophyll content had a limited reduction, however, and thevariance of yield was only reduced to 91%. When the field waspartitioned using LAI and leaf chlorophyll content, thevariances of LAI and leaf chlorophyll at a specific IFC hadsignificant reductions, whereas the variances of yield andelectrical conductivity had very limited reductions. Thevariance reduction of LAI and leaf chlorophyll is not verystable at IFC1. This is because the fraction of crop cover (corn)was very low at that time and therefore the estimated LAI has avery small dynamic range and the estimated leaf chlorophyllcontent is somewhat uncertain (Haboudane et al., 2002; 2004).For the crop-based delineation, the variances of LAI and leafchlorophyll content were reduced to about 22% and 67% atIFC2 and 26% and 89% at IFC3, respectively. Compared withfield 25, the variance of corn yield in field 23 is less accountedfor by soil features and crop descriptors.

Spatial patterns in field 25

Figure 4 shows the results of soil zone delineations in field25 using the two soil feature sets, e.g., BI and CI and PCs. Theresults of crop zone delineation using LAI and leaf chlorophyllcontent for IFC1, IFC2, and IFC3 are given in Figure 5. Forconvenience, the partitioned zones from the two soil featuresets are referred to as soil-based zones, and those from LAI andleaf chlorophyll content as crop-based zones. Soil-based zonesand crop-based zones are indicated by the subscripts s and v,which refer to soil and vegetation, respectively.

The soil zones delineated from BI and CI and PCs showsome similarities. The spatial patterns of the soil-based zonesgenerally match the soil type distribution as revealed by the soilsurvey map (Figure 1). C1s and C2s mostly represent D3 soilseries, and C4s mostly represents M3–NG2 soil series. C3sdistributed in between these two regions may be the transition

of these two soil associations. The patterns of C1s and C2sdelineated using BI and CI are not consistent with thedelineation using soil PCs. The soil properties may not be verydifferent between these two delineated zones. C1s delineatedusing BI and CI at the upper boundary of the field representsthe slope area with low organic matter, low sand content, andhigh clay content, which indicates soil erosion due to the slopeheading toward the creek flowing between the wheat (field 25)and corn (field 23) fields. This pattern is also well defined inthe crop zones at IFC1 and IFC2 stages and is typical of lowLAI and yield. Soil leveling and stabilization might be requiredin this area.

Different levels of N were applied with a specific pattern inthe wheat field (Figure 5d). The applied N was 0, 41, or 68 kgN ha–1 (referred to as 0N, 41N, and 68N, respectively). Thecrop-based zones derived from the partitioning of LAI and leafchlorophyll content show the combined effects of soilproperties and N application. The 0N application area at thesouthwestern corner is clearly delineated as C1v throughout theseason. Statistics show that this region has lower final yield,lower LAI, and lower leaf chlorophyll content at all the threeIFCs compared with the other zones. N deficiency is the majorcritical concern. Although recommended amounts of nitrogen(68N) were applied to the slope area, LAI in this region wassignificantly lower at IFC1 and IFC2 stages. This region, beingclearly delineated as C1v at IFC1 and IFC2 stages, is notfavourable for crop growth and led to a lower yield, which waspresumably caused by the lack of organic matter as a result ofsoil erosion toward the creek. Class C4v defined in IFC1 (upperright corner) overlaps with regions of the M3–NG2 soil series.In this zone, LAI and leaf chlorophyll content of the first twoIFCs and final yield have higher values, indicating that the soiltype in this area is favourable for crop growth. The cropdeveloped faster in this area than in the other areas, whichmakes wheat reaching its senescence stage earlier. Thedecrease of leaf chlorophyll content and green LAI in thisregion accounted for it being classified as C1v and C2v at theIFC3 stage.

Spatial patterns in field 23

Figure 6 shows the results of soil-based zone delineation infield 23 using the two soil feature sets. Results of crop-basedzone delineations for IFC1, IFC2, and IFC3 are given inFigure 7.

The similarity is weaker between the soil-based zonesresulting from BI and CI and PCs in field 23. C1s mostlyrepresents the poorly drained, fine-textured Brandon series(D3), C2s is mostly associated with poorly drained andimperfectly drained soils of Allendale and Montain seriesassociations (M6 and M3), and C3s represents a well-drained toimperfectly drained sandy soil association (U2–M5)regrouping deep (>100 cm) sandy soils (Carlsbad andRamsayville series) and shallow (25–100 cm) sandy soils overclay material (Allendale and Montain series).



Figure 4. Soil delineation of field 25 into four classes using (a)soil brightness and colouration indices, and (b) principalcomponents. C1s–C4s, soil-based zones 1–4, respectively.

Corn planted in field 23 received a uniform recommended Napplication. The spatial patterns of the crop classes aretherefore mostly caused by the interaction between soil andweather conditions. Crop-based delineation is difficult at IFC1because corn was at the early emergence stage in the field.Differentiation of C1v, C2v, and C3v at the IFC1 stage is mainlydue to the amount of vegetation cover and the soil properties.At IFC1, LAI in C3v generally ranges from 0.3 to 0.6, whereasin C1v and C2v it is less than 0.3. C3v delineated at the IFC1stage also has higher LAI values at IFC2 (from 2.5 to 4.0) and

IFC3 (>4.0) than the other delineated zones. The crop-basedzones are best delineated at the IFC2 stage, in that the variancesof both LAI and leaf chlorophyll content are significantlyreduced (Figure 3).

Statistical analysis

Relationship between the two soil feature setsSoil brightness dominates the spectral variance among soils.

The first principal component of soil data accounts for the



Figure 5. (a–c) Crop delineation of field 25 into four classes using LAI and leaf chlorophyll content at IFC1, IFC2,and IFC3, respectively. C1v–C4v, vegetation (crop) based zones 1–4, respectively. (d) Nitrogen application pattern(0N, 41N, and 68N denote applications of 0, 41, and 68 kg N ha–1, respectively).

Figure 6. Soil delineation of field 23 into three classes using (a) soil brightness andcolouration indices and (b) principal components.

Figure 7. Crop delineation of field 23 into three classes using LAI and leaf chlorophyll content at (a) IFC1, (b) IFC2,and (c) IFC3.

majority of the variability and represents approximately theaverage value of the spectrum, and therefore it is a measure ofsoil brightness. The first principal component and soilbrightness calculated using Equation (1) are highly linearlycorrelated, with determination coefficients (R2) of 0.998 and0.970 in fields 25 and 23, respectively. This explains thesimilar results of delineations using PCs and BI and CI(Figures 2, 3, 4, and 6). The different information contentbetween the higher order principal components (the second andthird) and the chromatic component CI mainly accounts for thedifference of the delineations.

Correlation between crop descriptors and yieldThe correlation between crop descriptors and yield was

analyzed, and the results are given in Table 2. In field 25, thecorrelations between LAI and wheat yield and between leafchlorophyll content and wheat yield are significant at IFC1 andIFC2 but not significant at IFC3. In field 23, yield is notsignificantly correlated with LAI or leaf chlorophyll content.This is consistent with the results shown in Figures 2 and 3: thevariance of yield in field 25 was reduced to 82% and 85% forthe crop-based delineation at IFC1 and IFC2, respectively, butthere was no significant reduction for the crop-baseddelineation at IFC3, and there was very limited variancereduction of corn yield in field 23 for the crop-baseddelineation at any of the three IFCs. The possible reason for thepoor correlation in field 23 is that LAI and leaf chlorophyllcontent did not capture a high productivity spatial featureacross the field, which decreased the overall correlation. If thishigh productivity feature is masked out, then a significantcorrelation is observed between corn yield and leaf chlorophyllcontent at IFC2 and IFC3 (Table 2, field 23A). The highestcorrelations were obtained with leaf chlorophyll content.

Zone means and analysis of varianceFigures 8 and 9 show the zone means and standard

deviations of the variables in the delineated zones of fields 25and 23, respectively. For variables EC30, EC100, and yield, thefigures show zone means and deviations of the fivedelineations: crop-based delineations at the three IFCs, andsoil-based delineations using BI and CI and PCs. For LAI andleaf chlorophyll content, zone means and deviations of cropdelineations at the three IFCs were illustrated. Tukey’s test wasapplied to test the difference of zone means of the variables toevaluate the uniqueness of the delineated zones. The results arealso shown in Figures 8 and 9. Zones in which the mean values

do not differ significantly at the 95% confidence interval aremarked with a box above the data bars. For instance, in field 25,the means of soil zones C1 and C2 delineated by BI and CI donot differ significantly at the 95% confidence interval. In thiscase, a box is shown above the data bar spanning C1 and C2(see Figure 8a).

In field 25, EC30, EC100, and yield differ significantlybetween soil-based zones except between C1 and C2. Thismeans that soil features extracted from hyperspectral datarevealed some of the soil properties, and the detected soilproperties had an important impact on the final yield. It can beobserved that yield in this field is negatively related to electricalconductivity. Yield is highest in the soil-based zone C4 andlowest in zones C1 and C2, and EC30 and EC100 are lowest inzone C4 and highest in zones C1 and C2. The high electricalconductivity corresponds to heavier soil texture, and these soilstend to stay saturated for longer periods of time, which isnegative for yield. For the crop-based zones, LAI and leafchlorophyll content differ significantly among the zones,whereas EC30 and EC100 do not differ significantly betweensome of the crop-based zone pairs. The crop-based zones atIFC3 do not effectively differentiate wheat yield, whereas theyare indicative of yield at IFC1 and IFC2. This means that theeffective time for delineation of the wheat crop should beearlier than that at IFC3.

In field 23, soil electrical conductivity differs significantlybetween all pairs of soil-based zones, whereas there is nosignificant difference in electrical conductivity between thecrop-based zones (Figure 9). Yield does not differ significantlyamong the soil-based zones as well as it does among the crop-based zones in this field. Except for LAI at IFC1 and leafchlorophyll content at IFC3, crop descriptors differsignificantly between all pairs of crop-based zones. Zone C1delineated at IFC1 has a high yield compared with that in theother zones because it is completely within a high-productionregion in the field. From IFC1 to IFC3, LAI in field 23increased steadily. For corn in field 23, the effective time fordelineation of crop-based zones should be later than that forIFC1.

ConclusionsIn this study, multitemporal CASI hyperspectral data were

used for zone delineation of two agricultural fields. Differentfeatures extracted from hyperspectral data related well to someof the field variables and revealed the variability of seasonallystable and variable information useful for management zonedelineation for precision agriculture.

The variability in soil electrical conductivity can beaccounted for to a significant extent by the features extractedfrom hyperspectral soil reflectance data. Several inherent soilfertility indicators like soil texture components (sand, silt, andclay content) and exchangeable cations (Ca and Mg) and soildrainage and related soil moisture conditions could be relatedto soil electrical conductivity. This is rarely the case, however,for the organic matter content of the surface layer, which is



Field LAI1 LAI2 LAI3 Chl1 Chl2 Chl3

25 0.57** 0.51** 0.23 0.54** 0.43* 0.1523 –0.09 –0.10 –0.08 –0.19 –0.13 0.2223A 0.04 0.15 0.22 –0.13 0.54** 0.68**

Note: The suffixes 1–3 to LAI and Chl represent IFC1–IFC3,respectively. *, significant at p < 0.01; **, significant at p < 0.001. Forfield 23A, the high productivity feature was masked out.

Table 2. Correlation coefficients between yield and cropdescriptors in the two fields.



Figure 9. Zone statistics and multiple comparisons for field 23.Variables include electrical conductivity at 0–30 cm (EC30) and 0–100 cm (EC100), yield, LAI, and chlorophyll content; delineationsinclude crop-based delineation at IFC1, IFC2, and IFC3 and soil-based delineations using BI and CI and PCs. The boxes above thedata bars indicate the classes that do not differ significantly at the95% confidence interval. The vertical bars denote standarddeviation.

Figure 8. Zone statistics and multiple comparisons for field 25.Variables include electrical conductivity at 0–30 cm (EC30) and 0–100 cm (EC100), yield, LAI, and leaf chlorophyll content;delineations include crop-based delineation at IFC1, IFC2, andIFC3 and soil-based delineations using BI and CI and PCs. Theboxes above the data bars indicate the classes that do not differsignificantly at the 95% confidence interval. The vertical barsdenote standard deviation.

most directly associated with soil reflectance. Therefore,remote sensing data have been shown to play a strong role insoil delineation and could be viewed, under given conditions,as an efficient alternative to soil conductivity mapping fordefining within-field homogeneous management zones.

The crop descriptors derived from hyperspectral data arevery useful for monitoring crop growth conditions. Theyrevealed the effects of soil properties under natural growthconditions and the effects of special nitrogen application undercontrolled conditions. The appropriate time to delineate wheatin field 25 was at IFC1 and IFC2 (prior to senescence), and theappropriate time to delineate corn in field 23 was after IFC1(after complete emergence). Crops can be monitored frequentlyusing LAI and leaf chlorophyll content to monitor seasonallyvariable information to guide the real-time field practices.

Zone delineation was evaluated by the variance reduction ofyield. From this perspective, field 25 is better delineatedbecause the variance of yield was reduced significantly for soiland crop delineations. The soil properties in this field have animportant impact on final yield, and crop descriptors LAI andleaf chlorophyll content at the earlier stages (IFC1 and IFC2)are indicative of final yield. Field 23 is not well delineated interms of variance reduction of yield.

In this study, delineation of management zones of the fieldsis based solely on the classification of the features extractedfrom hyperspectral data. Soils and crops were delineatedindependently using multitemporal hyperspectral data. Thecombination of soil features and crop descriptors beforedelineation, or the combination of the delineated results, couldgive better results for delineation of management zones. Theintegration of other sources of information, such as soilproperties, environmental conditions, and field managementfactors, may also greatly improve the quality and usefulness(i.e., interpretability) of management zone delineation. Inaddition to the acquisition of the information on fieldvariability, a more complete understanding of the causes ofcrop production variability is of great importance, since it willimprove the efficiency of the information integration formanagement zone delineation and its usefulness fordevelopment of management strategy. This can be achieved byperforming field delineation over several growing seasons tocapture the effects of several weather pattern incidences oncrop growth.

AcknowledgementsThis study was funded by the Government Related Initiatives

Program (GRIP) project between the Canadian Space Agencyand Agriculture and Agri-Food Canada and a GEOIDE(Canadian Networked Centre of Excellence in geomatics)project for CASI acquisition in 2000.

ReferencesBezdek, J. (Editor). 1981. Pattern recognition with fuzzy objective function

algorithm. Plenum Press, New York.

Doerge, T. 1998. Defining management zones for precision farming. CropInsights, Vol. 8, No. 21. Pioneer Hi-Bred International, Inc., Johnston,Iowa.

Duda, T., and Canty, M. 2002. Unsupervised classification of satelliteimagery: choosing a good algorithm. International Journal of RemoteSensing, Vol. 23, No. 11, pp. 2193–2212.

Escadafal, R., Belghith, A., and Ben Moussa, H. 1994. Indices spectraux pourla télédétection de la dégradation des milieux naturels en Tunisie aride. InProceedings of the 6th International Symposium on PhysicalMeasurements and Signatures in Remote Sensing, 17–21 January 1994, Vald’Isère, France. CNES. pp. 253–259.

Fridgen, J.J., Kitchen, N.R., and Sudduth, K.A. 2000. Variability of soil andlandscape attributes within sub-field management zones. In Proceedings ofthe 5th International Conference on Precision Agriculture, 16–19 July2000, Bloomington, Minn. CD-ROM. Edited by P.C. Robert, R.H. Rust,and W.E. Larson. ASA–CSSA–SSSA, Madison, Wis.

Haboudane, D., Miller, J.R., Tremblay, N., Zarco-Tejada, P.J., andDextraze, L. 2002. Integrated narrow-band vegetation indices for aprediction of crop chlorophyll content for application to precisionagriculture. Remote Sensing of Environment, Vol. 81, pp. 416–426.

Haboudane, D., Miller, J.R., Pattey, E., Zarco-Tejada, P.J., and Strachan, I.2004. Hyperspectral vegetation indices and novel algorithms for predictinggreen LAI of crop canopies: modeling and validation in the context ofprecision agriculture. Remote Sensing of Environment, Vol. 90, pp. 337–352.

Huete, A.R., and Escadafal, R. 1991. Assessment of biophysical soilproperties through spectral decomposition techniques. Remote Sensing ofEnvironment, Vol. 35, pp. 149–159.

Lelong, C.C.D., Pinet, P.C., and Poilve, H. 1998. Hyperspectral imaging andstress mapping in agriculture: a case study on wheat in Beauce (France).Remote Sensing of Environment, Vol. 66, pp. 179–191.

Leone, A.P., and Escadafal, R. 2001. Statistical analysis of soil colour andspectroradiometer data for hyperspectral remote sensing of soil properties(example in a southern Italy Mediterranean ecosystem). InternationalJournal of Remote Sensing, Vol. 22, No. 12, pp. 2311–2328.

Liu, J.G., and Moore, J.Mcm. 1990. Hue image RGB colour composition: asimple technique to suppress shadow and enhance spectral signature.International Journal of Remote Sensing, Vol. 11, No. 8, pp. 1521–1530.

Marshall, I.B., Dumanski, J., Huffman, E.C., and Lajoie, P. 1979. Soilscapability and land use in the Ottawa urban fringe. Land ResourceResearch Institute, Agriculture Canada, Ottawa, Ont. 59 pp.

Mattikalli, N.M. 1997. Soil colour modelling for the visible and near infraredbands of Landsat sensors using laboratory spectral measurements. RemoteSensing of Environment, Vol. 59, pp. 14–28.

Moran, M.S., Inoue, Y., and Barnes, E.M. 1997. Opportunities and limitationsfor image-based remote sensing in precision crop management. RemoteSensing of Environment, Vol. 61, pp. 319–346.

Nolin, M.C., Gagnon, B., Leclerc, M.-L., Cambouris, A.N., Belanger, G., andSimard, R.R. 2002. Influence of pedodiversity and land uses on the within-field spatial variability of selected soil and forage quality indicators. InProceedings of the 6th International Conference on Precision Agriculture,14–17 July 2002, Minneapolis, Minn. CD-ROM. Edited by P.C. Robert etal. ASA–CSSA–SSSA, Madison, Wis.

Pattey, E., Strachan, I.B., Boisvert, J.B., Desjardins, R.L., and McLaughlin,N.B. 2001. Detecting effects of nitrogen rate and weather on corn growth



using micrometeorological and hyperspectral reflectance measurements.Agriculture and Forest Meteorology, Vol. 108, pp. 85–99.

PCI Geomatics Enterprises Inc. 2001. EASI/PACE software, version 8.2.3. PCIGeomatics Enterprises Inc., Richmond Hill, Ont.

Perron, I., Cluis, D.A., Nolin, M.C., and Leclerc, M.-L. 2002. Influence ofmicrotopography and soil electrical conductivity on soil quality and cropyields. In Proceedings of the 6th International Conference on PrecisionAgriculture, 14–17 July 2002, Minneapolis, Minn. CD-ROM. Edited byP.C. Robert et al. ASA–CSSA–SSSA, Madison, Wis.

Perron, I., Nolin, M.C., Pattey, E., Bugden, J.L., and Smith, A. 2003.Comparaison de l’utilisation de la conductivite électrique apparente (CEa)des sols et des donnees polarimetriques RSO pour delimeter des unitesd’amenagement agricole. In 25th Canadian Symposium on RemoteSensing – 11e Congrès de l’Association québécoise de télédétection, 14–16October 2003, Montréal, Que. CD-ROM.

Rafter, J.A., Abell, M.L., and Braselton, J.P. 2002. Multiple comparisonmethods for means. SIAM Review, Vol. 44, No. 2, pp. 259–278.

Shibayama, M., and Akiyama, T. 1991. Estimating grain yield of maturing ricecanopies using high spectral resolution reflectance measurements. RemoteSensing of Environment, Vol. 36, pp. 45–53.

Strachan, I.B., Pattey, E., and Boisvert, J.B. 2002. Impact of nitrogen andenvironmental conditions on corn as detected by hyperspectral reflectance.Remote Sensing of Environment, Vol. 80, pp. 213–224.

Thenkabail, P.S., Smith, R.B., and De Pauw, E. 2000. Hyperspectral vegetationindices and their relationships with agricultural crop characteristics.Remote Sensing of Environment, Vol. 71, pp. 158–182.

Tomer, M.D., Anderson, J.L., and Lamb, J.A. 1995. Landscape analysis of soiland crop data using regression. In Proceedings of Site-specific Managementfor Agriculture Systems, 27–30 March 1994, Minneapolis, Minn. Edited byP.C. Robert, R.H. Rust, and W.E. Larson. ASA–CSSA–SSSA, Madison,Wis. pp. 273–284.

Tucker, C.J., Holben, B.N., Elgin, J.H., Jr., and McMurtrey, J.E. 1980.Relationship of spectral data to grain yield variations. PhotogrammetricEngineering and Remote Sensing, Vol. 46, pp. 657–666.

Zhang, N., Wang, M., and Wang, N. 2002. Precision agriculture — aworldwide overview. Computers and Electronics in Agriculture, Vol. 36,pp. 113–132.



Documents

Can. J. Remote Sensing, Vol. 31, No. 5, pp. 400–411, 2005